The Web3 + AI Daily: BASS 2026
Discussions on NEAR AI, Gensyn, ChainOpera, Bittensor, and Pluralis Research from this year's Blockchain Application Stanford Summit.
Hello, everyone, and welcome to all newly joined subscribers!
This is the 78th edition of The Web3 + AI Daily - your definitive guide to the intersection of blockchain and AI. After highlighting the most interesting Web3 + AI talks from this year’s Science of Blockchain Conference, I decided to do the same with one of SBC’s major side events - the Blockchain Application Stanford Summit (BASS).
While the SBC showcases the latest research and innovation across cryptography and blockchains, BASS centers on Web3’s application layer. This year’s agenda reflects a notable shift: after years of blockchain applications barely making it into live production, now a growing number of dApps are up and running, with blockchains increasingly underpinning real-world financial activity.
BASS is organized by Gil Rosen, 🌊Steven Willinger 🌊, and Kun P., also running the Blockchain Builders Fund and the Stanford Blockchain Accelerator, and sponsored by 1inch, Cambrian Network, PrismaX, 0G, TRON DAO, ChainOpera AI, and others.
Thank you for being here! Let’s dive in.
NEAR’s Illia Polosukhin on Sovereign AI
NEAR Protocol and NEAR AI's Illia Polosukhin took the BASS stage to examine a palpable shift in AI: users no longer care about AI’s capabilities only, but instead start to question its long-term effects on businesses, the global economy, and even democracy. This is where Web3’s emphasis on private and sovereign AI naturally begins to resonate.
Lately, the conversation around AI has been revolving around closed vs. open-source, but having access to a model’s open weights is not enough. Rather, an entire technology stack is required to ensure a truly safe and sovereign AI. It spans:
Secure hardware
Verifiable model and training data
Execution
Inference
Coordination and discovery
Agent / harness
NEAR has been working on several parts of the stack, including confidential inference, where blockchain plays an important role. There are two aspects of confidential AI:
Verifiability: the ability to check both what model is working in the background and what data it was trained on.
“Without having verifiability, you’re at the mercy of whoever has the hardware to affect how you use AI”
Privacy: making sure that your corporate and personal data remain encrypted and is not being used for training by model providers.
Polosukhin used the occasion to share an intriguing announcement:
“You can now stake $NEAR tokens and continuously receive credits for confidential inference on NEAR AI or deploy IronClaw agents.”
This launch enables $NEAR stakers to receive monthly compute credits proportional to their stake size. It changes the relationship between owning NEAR’s token and accessing AI products, directly tying tokenomics to AI inference usage.
Moreover, it mirrors the inference tokenization model employed by Venice.ai, where $VVV stakers can mint $DIEM and access daily inference credits (learn more here). A secondary inference market has already emerged around $DIEM, let’s see if the same will happen with NEAR.
Listen to the full presentation here.
Decentralized AI Inference and Training
Gensyn's Ben Fielding, ChainOpera AI's Salman Avestimehr, and Ala S., co-founder of Bittensor / Opentensor Foundation and Crucible Labs, joined a panel on decentralized AI inference and training. The two key takeaways I took from this discussion are:
Producing decentralized AI comes with serious tradeoffs, so the dAI space is gradually coming to terms with the fact that not every AI use case must be decentralized.
Accordingly, dAI should concentrate on the areas where it is uniquely positioned to deliver value.
And no - nobody’s saying that dAI cannot compete with centralized AI. Quite the opposite: as Fielding pointed out:
“In the long term, more powerful models can be made through decentralized means, but we have to overcome a significant number of hurdles to get there. What the entire decentralized AI space has realized is that the extra cost and the extra friction of doing something in a decentralized way mean that the only viable way to do it is to be selling something that is only achievable via decentralized AI.”
With the AI landscape evolving at a remarkable pace, dAI should stop playing catch-up with AI hyperscalers, and rather focus on leveraging its unique strengths. Fortunately, the sector seems to have matured enough to realize that.
Fielding gave an example of just that. Back in 2019 and 2020, compute (or the lack thereof) was considered the biggest bottleneck in AI, which led Gensyn to explore decentralized AI training and coordination of distributed GPUs. Yet, proving that a particular GPU exists and it’s owned by a specific miner, is not an easy problem. Doing it is actually more costly than relying on a centralized cloud, so the time, money, and efforts spent there might be better invested somewhere else.
This has been the story of dAI over the past few years: a process of trial and error that has gradually led to a much narrower focus.
Bittensor is a compelling illustration of an AI product that only a blockchain could power. There have been projects aimed at coordinating distributed computing resources before, such as @SETI at home, but participants there had no incentive beyond the excitement of the possible discovery.
Bittensor changed that by introducing an incentive mechanism that rewards honest participation and penalizes attempts to game the system. As a result, compute resources were no longer just voluntarily lent, but exchanged in a fully-functioning decentralized compute market.
In that sense, the blockchain network isn’t just an alternative infrastructure layer. It’s what makes the market itself possible.
The panel concluded with a discussion of developments that could potentially affect dAI, both positively and negatively. The resurgence of open source and the growing trend toward owning your software were cited as positive trends. On the other hand, overregulation, to the extent where open-source AI is banned, as well as a potential bursting of the AI bubble, were identified as risks that could have repercussions far beyond dAI and spill over into the broader financial system.
Listen to the full panel recording here.
Decentralized Training: Unattainable Dream or Viable Alternative
If the previous talk made you believe that decentralized AI training is a lost cause, this one will convince you otherwise. Alexander Long, founder of Pluralis Research, sat down for a fireside chat with CoinFund‘s Jake Brukhman to explore how Pluralis is tackling the challenges of decentralized AI training.
Let’s start form the very beginning - what is decentralized AI training? As Brukhman commented:
“AI needs giant data centers with very high-end compute, with very very high bandwidth links between the GPUs, and a lot of energy to create large frontier models. It’s getting to the point that some companies are buying nuclear power plants to power these things. Decentralized training is the idea that we can do the same type of training done in these data centers, but in a swarm of consumer hardware.”
A few years ago, it was largely believed that training AI over a distributed swarm of consumer devices was impossible. Now, there are two different approaches to achieving it:
Federated learning: Every independent node holds a full replica of the model, trains it locally on private data streams, and periodically sync updates across the network. Here, although data is broken down, hardware requirements still grow proportionately with the size of the model. So, there is a limit to how big a model could be trained that way.
Parallel training: Not just the data is broken down, but the model itself is also split over the participating nodes. That way, very large models can be trained on small nodes, which means that the hardware requirements are unbounded from the size of the model. This is what Pluralis is doing.
Almost all training methods rely on very high-bandwidth links between the GPUs. How Pluralis solves this is not by modifying the existing models, but by creating new model architectures designed specifically for a decentralized setting. This is not a new idea - optimizing AI models in accordance with the hardware used for training has been happening since the dawn of deep learning.
“The model Pluralis created is very similar to the Transformers, but where there’s massive compression on the output of the blocks, which allow for a transmission over much lower bandwidth links, which in turn allows for different people to operate different parts of the blocks.”
The focal point of this conversation for me was this: coming up with new and more efficient ways to perform decentralized AI training is a matter of continuous research, but what’s more important is the bigger picture and the ripple effect it could produce.
For Long, the rationale is quite clear:
“To me the motivation for decentralized training or protocol learning is: can you get a truly neutral substrate here, where it’s a bedrock that you can build on. And there’s a community development happening on top, no one can turn this thing off, no one can yank it up from under you, it’s just stable. Very analogous to: do you want the Linux of the models vs. the Windows. This happened at every layer of technology ever, and it hasn’t happened yet at the model layer, and to me decentralized training is the answer to that question.”
There is another key property of the Pluralis approach to decentralized training that deserves more attention:
Because the model is split across participants, no single party owns the model weights. As a result, the traditional business model of licensing access to open weights doesn’t work. Instead, Pluralis monetizes the exclusivity inference of the model and programmatically achieves what, for instance, Kimi (Moonshot AI) achieves through licensing.
“You could theoretically create an open model in a decentralized swarm, and give away the weights. That’s definitely an option. But the more interesting option is to build the same business model that OpenAI and Anthropic have, but you have it collectively and publicly owned.”
Finally, none of this would be possible without blockchain technology, which addresses both the capital and ownership questions. There’s no sustainable way to bring such enormous amounts of compute into the protocol without providing economic incentives and allowing participants to claim ownership of the resulting model.
Listen to the full talk here.
Thank you for reading! I hope you found this article insightful.
If you’d like to keep up with the latest developments at the burgeoning intersection of blockchain and AI, I invite you to subscribe.
And if you find value in my writing, consider supporting the newsletter. I’d greatly appreciate it.
The Web3 + AI Book Club is live on Fable. We’re currently reading Karen Hao’s ‘Empire of AI.’ Follow the link below and come join our literary community.
I'm looking forward to connecting with fellow Crypto x AI enthusiasts, so don't hesitate to share your thoughts here or reach out on LinkedIn, Instagram, or Firefly.social.
Disclaimer: The content of this newsletter is provided for informational and educational purposes only. Nothing contained herein should be construed as financial advice or as a recommendation to buy, sell, or hold any of the companies or assets mentioned.
Please don’t take my views at face value. Instead, do your own research (DYOR), think critically, and share your perspective so we can challenge ideas, learn from one another, and arrive at better conclusions through thoughtful discussion.